Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: goodness of fit measure fir ivtobit


From   Nick Cox <njcoxstata@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: goodness of fit measure fir ivtobit
Date   Mon, 24 Sep 2012 17:55:34 +0100

Agreed; but the opposite of the problem here, which is that low
R-squared may represent a worthwhile attempt to find some pattern in
highly variable data.

Nick

On Mon, Sep 24, 2012 at 5:47 PM, Yuval Arbel <yuval.arbel@gmail.com> wrote:
> regarding my last quote: we are talking about annual data between 1897
> and 1958. I guess the authors' objective was to demonstrate that high
> correlation could be spurious
>
> On Mon, Sep 24, 2012 at 6:43 PM, Yuval Arbel <yuval.arbel@gmail.com> wrote:
>> Nick, if we are aware of the many limitations of the R-Squares that's
>> O.K. to report it. My fear is that during a presentation somebody will
>> see the low R-Squared and say that given that 88 percent of the
>> variance cannot be explained  Anat's work is worthless, where this is
>> obviously not the case (and believe me, these things happened before).
>>
>> Finally, a nice anecdote from Johnston and Dinardo's textbook
>> (Econometric Methods fourth edition (1997) page 10): they quote
>> Plosser and Schwert, who found a +0.91 correlation between the log of
>> U.S. nominal income and the log of accumulated sunspots!!!
>>
>> On Mon, Sep 24, 2012 at 6:26 PM, Nick Cox <njcoxstata@gmail.com> wrote:
>>> That is a puzzling argument, if indeed it's an argument at all.
>>>
>>> I don't think anything much in statistics ensures a _causal_
>>> relationship  (presumably what Yuval means here), short of independent
>>> evidence on mechanism or process.
>>>
>>> If a model is not that great, readers need to know. Sometimes a low
>>> R-squared makes that vivid.
>>>
>>> (People who want to remind me how limited R-squared is should please
>>> note that I wrote the FAQ cited below, which comes decorated with
>>> multiple warnings.)
>>>
>>> Nick
>>>
>>> On Mon, Sep 24, 2012 at 5:12 PM, Yuval Arbel <yuval.arbel@gmail.com> wrote:
>>>> Anat, note that the possibility to calculate the log likelihood is
>>>> there regardless of the method of estimation you are employing.
>>>>
>>>> In addition, I would personally rather avoid presenting an R-Squared
>>>> of 0.12, particularly in these kinds of models. As is well known, high
>>>> R-Squared does not ensure casual relationship and low R-Squared does
>>>> not ensure lack of casual relationship
>>>>
>>>> On Mon, Sep 24, 2012 at 4:54 PM, Anat (Manes) Tchetchik
>>>> <anatmanes@gmail.com> wrote:
>>>>> I haven't thought about the count model, I will definitely try to run
>>>>> it! thanks much!
>>>>>
>>>>> On Mon, Sep 24, 2012 at 5:38 PM, Maarten Buis <maartenlbuis@gmail.com> wrote:
>>>>>> That does not sound like censoring at all. I would think of this as a
>>>>>> regular count model. There are examples on how to deal with such an
>>>>>> iv-model in -help gmm-.
>>>>>>
>>>>>> Hope this helps,
>>>>>> Maarten
>>>>>>
>>>>>> On Mon, Sep 24, 2012 at 4:11 PM, Anat (Manes) Tchetchik
>>>>>> <anatmanes@gmail.com> wrote:
>>>>>>> Austin Hi,
>>>>>>> Thank you very much for your reply!
>>>>>>> What I have as a dependent var. are 500 respondents' reports of the
>>>>>>> number of times they travelled abroad to visit their friends and
>>>>>>> relatives over the course of their adult lives.  Some respondents yet,
>>>>>>> who have relatives abroad, did not travel at all.
>>>>>>> So the observations are censored at zero, with mean =2.2, max =50 and
>>>>>>> stdev= 3.8.
>>>>>>> Do you think in that case that the general methods of moments will be better?
>>>>>>> Thanks much!!!
>>>>>>> Anat
>>>>>>>
>>>>>>> On Sun, Sep 23, 2012 at 5:49 AM, Austin Nichols <austinnichols@gmail.com> wrote:
>>>>>>>>
>>>>>>>> Anat (Manes) Tchetchik <anatmanes@gmail.com>:
>>>>>>>> You can always -predict- and compute the squared correlation of
>>>>>>>> predictions with observed values:
>>>>>>>> http://www.stata.com/support/faqs/statistics/r-squared/
>>>>>>>> but are you sure your -ivtobit- model is justified?  What is the
>>>>>>>> process that results in observations being censored?  I suspect you
>>>>>>>> have a lower limit at zero which is actually a very low conditional
>>>>>>>> mean rounded down to zero--am I right?  You may be better off with a
>>>>>>>> -gmm- model.
>>>>>>>>
>>>>>>>> On Sat, Sep 22, 2012 at 5:33 PM, Anat (Manes) Tchetchik
>>>>>>>> <anatmanes@gmail.com> wrote:
>>>>>>>> > Dear statalisters,
>>>>>>>> >
>>>>>>>> > I wonder if anyone knows any goodness of fit that is appropriate for
>>>>>>>> > tobit with endogenous
>>>>>>>> > variables (ivtobit). Not as in "regular" tobit, stata does not report any
>>>>>>>> > goodness of fit measure, any idea how to estimate such a measure?
>>>>>>>> > Any response will be greatly appreciated..
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index